異常行為事件之偵測在視覺監視系統中係一項重要的研究課題,尤其在擁擠環境之影像中,準確且可信賴的追蹤和偵測往往是一項重大挑戰。本論文將考量連續影像中時間和空間的特徵關係,基於社會力模型(Social Force Model)的概念而提出HOSF方法作為特徵描述,來解決傳統基於物件路徑之偵測方法不適用於擁擠影像的問題。經由粒子的分割建立cuboids,再透過HOSF特徵向量訓練出判斷事件異常性所依據的字典(dictionary),最後根據z-value計算事件與字典中codeword的相似情形來決定影像中事件的異常與否。 本論文所提方法包含以下特點:(1)訓練過程為全自動,不需事先經由人工決定事件屬於正常或異常;(2)特徵描述採用粒子和社會力,不需事先對物件做追蹤,因此不論擁擠人群或人少的場景皆可適用;(3)使用z-value方法來估算事件的異常性,由於計算簡單,訓練完成後即可達到即時偵測。 In this paper a simple and effective crowd behavior normality method is proposed. Feature vector, so called HOSF (histogram of oriented social force), and consists of concatenating local histogram of oriented social force. A dictionary of codewords is trained to include typical HOSF. To detect whether an event is normal is accomplished by comparing how similar to the closest codeword via z-value. The proposed method includes the following characteristic: (1) the training is automatic instead of human labeling; (2) instead of object tracking, the method integrates particles and social force as feature descriptors which well adapted in both crowded or few people scenes; (3) z-score is used in measuring the normality of events. Due to computation simplicity, the normality detection could be real-time once the training is finished.